Introduction

DeepSeek Harness (dsh) desktop natively supports multimodality (image understanding), but the official settings page does not provide an option to declare the input modality (input) for custom providers or third-party models. If an added model actually supports images, sending an image may still produce an error indicating “the current model does not support images.” The current workaround is to manually edit settings.yaml and add input: [text, image], which is cumbersome and error-prone. The dsh-vision-toggle plugin provides a graphical solution for this issue.

What Is This

This is a client-side plugin maintained by lijian-ui. It provides per-model “Support Image” toggles for DeepSeek Harness (dsh), allowing users to enable or disable image input with one click from the settings page, without manually editing configuration files.

Core Features

  • Model-level vision input toggles: Provides independent image input control for each configured model.
  • Reads routed models: Automatically reads all models under the llm-pi-ai provider route.
  • Immediate effect: Changes take effect immediately after toggling, without restarting the desktop application.
  • Optimistic UI feedback: Interface operations respond instantly, and automatically roll back if writing fails.
  • Internationalization support: Supports switching between Chinese and English interfaces.

Installation and Enablement

  1. Ensure the environment meets Node.js >= 18.
  2. Run the installation command:
    dsh plugin add @lijian-ui/dsh-vision-toggle
  1. After installation, restart the desktop application.

Typical Usage

  1. Open the dsh desktop application.
  2. Navigate to Settings -> Model Vision.
  3. This section lists all configured models (grouped by provider). Each model has a “Support image” toggle:
    • Enabled: The model accepts image input, and images can now be attached in the chat.
    • Disabled: The model switches to text-only mode.

Applicable Scenarios and Notes

  • Applicable scenarios: Mainly used for models managed by the llm-pi-ai provider route, especially custom or third-party models that actually support images but do not declare the input modality in the configuration.
  • Notes: The plugin only works under the llm-pi-ai provider route. After installation, the desktop application must be restarted to load the plugin. If writing the configuration fails, the toggle automatically reverts to its previous state.